[Pl-seminar] Semantics seminar schedule [Nonstandard time!]

[Aaron Turon]

NU Programming Languages Seminar presents

Vikash Mansinghka
Navia Systems, Inc.
MIT, Computational Cognitive Science Group

**Monday**, October 5, 2009
11:30-1:30
Room 366 WVH (http://www.ccs.neu.edu/home/wand/directions.html)

TITLE: Natively Probabilistic Computation

ABSTRACT:

Probabilistic inference, Bayesian reasoning and inductive learning are
at the heart of modern artificial intelligence and computational
science, but are often viewed as computationally intractable,
mathematically inaccessible, and difficult to scale to rich systems of
knowledge. I argue that these difficulties stem from a basic mismatch
between Bayesian probability and our deductive, deterministic view of
computation. Accordingly, I have been developing a stack of
abstractions for natively probabilistic computation, based on
distributions and samplers rather than functions and evaluators at
every level, from circuits to programs.

In this talk, I will focus on Church, a universal probabilistic
programming language for generative models, and Monte, a massively
parallel implementation of Church being developed at Navia Systems. I
will walk through several examples illustrating useful probabilistic
programs and develop connections between probabilistic and
deterministic programming, including generalizations of the notion of
purity that clarify both probabilistic and deterministic mutation. I
will also show how marrying probability with a reflective, dynamic
language like Lisp, including the ability to learn probabilistic
programs from data using general purpose inference machinery and the
ability to reason about the behavior of probabilistically
Turing-universal reasoners. Throughout, I will highlight opportunities
for collaboration in programming languages, compilers and
computational complexity.

If time permits, I will also touch on the rest of the probabilistic
computing stack, including probabilistic microarchitectures and
probabilistic gates. Our preliminary results show these layers can
yield 2-3 orders of magnitude improvements in price/power/performance
envelope and 3-6 orders of magnitude improvements in bit error
robustness over traditional deterministic microarchitectures and
Boolean circuits on probabilistic inference problems. Additionally,
these machines begin to address issues of fine-grained parallelism and
stochasticity in the deep submicron regime.

This talk includes joint work with Noah Goodman, Daniel Roy, Eric
Jonas, Keith Bonawitz, Beau Cronin and Joshua Tenenbaum.

BIO:

Vikash Mansinghka tries to make computers and programs that can manage
uncertainty and ambiguity as naturally as current computers can manage
arithmetic and deduction, by embracing randomness at every level of
abstraction. He is a co-founder of Navia Systems, Inc., a startup
company that is developing a commercial platform for probabilistic
computing and using this platform to help companies make sense of
their data. He is also a member of the Computational Cognitive Science
Group at MIT's Computer Science & Artificial Intelligence Laboratory
and Brain & Cognitive Sciences Department. He received S.B. degrees in
Mathematics and in Computer Science, an M.Eng. in Computer Science,
and a Ph.D in Computation, all from MIT, where he was a National
Science Foundation graduate fellow, a Lincoln Laboratory fellow, and
received the 2009 George M. Sprowls award for best dissertation in
computer science.

================================================================

If you are interested in giving a talk at the seminar, or know someone
who might be, please email turon at ccs.neu.edu